acceptodds
Withdrawn from ICLR 2027

Canonical Output Conditioning: Decoupling Navigation Interfaces from Diffusion Priors

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

A diffusion prior and the interface through which external conditions enter it are usually learned together: condition encoders, control branches, or classifier-free channels bind the trained model to the formats seen during training. Training-free guidance avoids retraining, yet still requires a compatibility function that relates each condition to the generated variable, typically a measurement operator, a differentiable loss, or a separately trained predictor. When condition and output live in different coordinates, constructing that function becomes its own learned component. We argue that this gap is fixed by a design choice made before either training or guidance: the choice of what the model generates. Canonical-output conditioning selects a representation whose coordinates both support the downstream task and admit direct comparison with externally converted conditions. The prior is then trained without conditions; heterogeneous formats reduce to explicit geometric adapters that feed a shared quadratic compatibility interface. We validate the design in route-guided navigation, where conditions are naturally heterogeneous. A LiDAR-conditioned diffusion prior over orientation fields, trained with no route inputs, is guided at inference by waypoint, corridor, and map route cues through Orientation Mixing, a confidence-weighted clean-estimate update. On six SemanticKITTI sequences, the single frozen prior serves all four interfaces and outperforms route-conditioned predictors trained on a fixed interface (mean ADE 0.29,m vs. 0.38,m), supporting representation choice as a practical lever for decoupling condition interfaces from diffusion priors.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Diffusion, Navigation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_uDBiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes canonical-output conditioning: external cues are expressed in the coordinates of a generated task representation and applied through a shared quadratic energy. GOF trains a LiDAR-conditioned diffusion prior over orientation fields without route inputs. Geometric adapters handle four route interfaces, and inference-time guidance comes from Orientation Mixing, re-noising, and spatial/history sectors. On six SemanticKITTI sequences, the paper shows fixed-prior interface reuse and reports 0.29 m mean ADE vs 0.38 m for deterministic OrField in the full system. It also runs intent, readout, temporal-conditioning, and map-perturbation experiments.

Critical strengths

I think the interface-reuse claim is well supported. Table 1 keeps GOF’s prior and four-sector layout fixed, swaps route interfaces, and drops temporal history. The paired intent study also controls observations and seeds: GOF follows the changed intent in 1022/1024 pairs vs 911/1024 for Conditional Training. That’s good evidence for field-level control without a trained route-input channel. The component studies are useful too. Re-noising cuts mean ADE from 0.44 to 0.31 m on all six sequences, and the readout-crossover and conflicting-history controls help clarify what trajectory extraction and temporal conditioning actually contribute.

Critical weaknesses

The accuracy gain from the conditioning architecture is still unclear. With the same greedy readout, the headline 0.38 vs 0.29 m becomes 0.31 vs 0.29 m, and both hit 96% HitRate (Table 9). GOF also has a richer cue/history setup, and the matched smoothed single-route result is missing. So the paper supports interface reuse more than a matched accuracy advantage.

The methodological step is also modest: the readouts come from OrField, Orientation Mixing is basically a gradient step on a quadratic energy, and backward Universal Guidance already does clean-estimate guidance. The DPS baseline is also underspecified, so its large deficit is hard to judge.

Questions for authors
  1. What does GOF get with the smoothed single-route input, no history, and a shared readout, compared with the learned-route models in their matched setting?
  2. How does GOF compare with backward Universal Guidance using the same field prior, quadratic energy, route cues, and denoiser budget?
  3. For DPS, what objective, differentiation variable, guidance scale, tuning protocol, and actual denoiser-evaluation count were used?
Reviewer_7VRDOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents Canonical-output Conditioning, a framework designed to decouple condition interfaces from diffusion priors through a shared output representation. The proposed GOF framework learns a LiDAR-conditioned diffusion prior over orientation fields without route inputs during training. At inference, different route conditions, including waypoints, corridors, single routes, and multiple routes, are incorporated through geometric adapters and Orientation Mixing. Experiments on SemanticKITTI evaluate navigation performance, route-interface flexibility, and the effects of different guidance components.

Critical strengths
  1. The paper is generally readable, and the overall framework is relatively easy to follow.
  2. The experiments consider several route representations and include component-level analyses.
Critical weaknesses
  1. The motivation and methodological novelty are not sufficiently established, particularly in relation to existing training-free guidance methods and orientation-field representations.
  2. The theoretical analysis is relatively straightforward and provides limited support for the broader claims of the proposed framework.
  3. The experimental comparisons contain potential confounding factors, and the evaluation is limited to a specific offline navigation setting.
Questions for authors
  1. The introduction argues that existing training-free guidance methods still suffer from difficulties in constructing compatibility functions. However, these methods can already incorporate external constraints through deterministic transformations. What specific limitation remains unresolved, and why is canonical-output conditioning necessary?
  2. The proposed framework adopts the orientation-field representation from prior work. What makes this representation canonical rather than simply a task-specific intermediate representation?
  3. Orientation Mixing is formulated as a weighted interpolation between the predicted clean field and external guidance. How does this operation differ fundamentally from existing output-space guidance or gradient-based conditioning methods?
  4. The theoretical analysis mainly establishes equivalence properties of quadratic energies and their local updates. How do these results support the claimed decoupling beyond properties that follow directly from the energy formulation?
  5. In Table 1, the learned baselines are evaluated under route-interface mismatch, while GOF can combine multiple route representations. To what extent can the performance differences be attributed to the proposed conditioning strategy rather than differences in the available route information?
  6. Table 2 reports an ADE improvement from 0.38 to 0.29, whereas Table 9 shows a smaller improvement from 0.31 to 0.29 under the same greedy readout, with identical HitRate. How should the contribution of the proposed method to the reported performance gains be interpreted?
  7. All four considered interfaces represent closely related geometric route information, and the evaluation is conducted on a single dataset. To what extent do these results support the generality of canonical-output conditioning beyond this specific representation and task?
  8. The evaluation uses offline route assignment based on recorded trajectories, without closed-loop execution. What conclusions about practical navigation can be drawn from the reported open-loop results?
Reviewer_cisgOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes choosing a generated representation that can serve as a shared coordinate system for both a learned prior and external conditions. Its GOF system trains a LiDAR-conditioned diffusion prior over orientation fields without route inputs, then applies waypoint, corridor, and route cues during inference through explicit adapters and Orientation Mixing. On six SemanticKITTI test sequences, it reports lower ADE/FDE and higher HitRate than the evaluated systems.

Critical strengths
  1. The paper gives a precise quadratic interface formulation and explains what an adapter must preserve for two conditions to induce the same update. This makes the proposed decoupling more than an informal architectural claim.
  2. The experiments include route-interface comparisons, component ablations, and tests of intent response and map perturbation. These offer targeted evidence for the design within the chosen setting.
Critical weaknesses
  1. The evaluation is offline and open-loop, uses one dataset, and does not test closed-loop behavior or safety. The paper acknowledges that these broader navigation properties remain untested.
  2. Route intent is assigned offline using alignment to logged trajectories, and key interface comparisons are reported on a single test sequence. This limits the evidence that the approach works when route inputs and intent must be obtained under deployment conditions.
Questions for authors
  1. At test time, how is the intended branch or local target sink assigned for each route interface? In particular, does the route-assignment procedure use the logged future trajectory, and if so, which parts of the results depend on that information?
  2. How do the primary navigation results vary across the six test sequences individually, especially for the route-interface comparison centered on sequence 08?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.